Overview
Agricultural yield statistics are typically reported at coarse administrative scales, while modern geospatial machine learning methods demand spatially explicit, pixel-level supervision. This mismatch has historically constrained the development of large-scale benchmarks for crop yield learning from multimodal Earth observation data.
To bridge this gap, researchers have introduced AgroBench, a reproducible benchmark that transforms publicly available U.S. county-level crop yield statistics into weakly supervised pixel-level crop time series. Each crop pixel time series is paired with a county-level yield value, which serves as a weak supervisory signal rather than a directly measured pixel-level yield label. This design enables scalable learning from readily available statistics without requiring costly field-level annotation.
Data Generation Pipeline
The AgroBench pipeline integrates multiple heterogeneous data sources into temporally aligned multimodal sequences that describe individual crop pixels throughout the growing season:
- USDA crop yield statistics β county-level yield values used as weak labels
- Crop-specific land cover masks β to isolate pixels belonging to each crop type
- Sentinel-2 multispectral imagery β optical observations across spectral bands
- Sentinel-1 synthetic aperture radar (SAR) β all-weather radar observations
- Climatic variables β temperature, precipitation, and related meteorological signals
- Terrain information β elevation and topographic features
Together, these modalities produce rich, time-aligned representations of crop growth dynamics at the pixel level.
Benchmark Scale
The resulting benchmark comprises:
- Over 13 million observations
- 788,654 unique crop pixels
- 5,107 county-year combinations
- Eight growing seasons (2017β2024)
- Five major U.S. crops
This scale makes AgroBench one of the largest weakly supervised multimodal datasets for agricultural yield prediction.
Evaluation Protocol and Baselines
To facilitate standardized and reproducible evaluation, the authors establish a crop yield prediction benchmark using a Leave-One-Year-Out (LOYO) evaluation protocol. This design tests a model's ability to generalize to unseen growing seasons, which is critical for real-world deployment under shifting climate and management conditions. Baseline results are provided using representative machine learning models, offering reference points for future comparisons.
Significance and 2026 Context
As of 2026, the convergence of geospatial foundation models, multimodal remote sensing, and weakly supervised learning has accelerated interest in agriculture-focused AI. AgroBench addresses a persistent bottleneck in this field β the absence of large, reproducible, and multimodal benchmarks β by releasing:
- The complete data generation pipeline
- The benchmark dataset
- The evaluation protocol
This release provides a reproducible foundation for future research in weakly supervised learning, multimodal remote sensing, spatiotemporal modeling, and geospatial foundation models for agriculture. By aligning coarse administrative statistics with fine-grained pixel observations, AgroBench enables the community to develop and rigorously compare methods that learn from weak, real-world supervision.
Paper Details
- Title: AgroBench: A Reproducible Multimodal Benchmark for Weakly Supervised Crop Yield Learning from County Statistics and Pixel Observations
- Authors: Udaiveer Singh, Rajiv Ranjan, Shashank Tamaskar, Dharmendra Saraswat
- Submitted: 16 September 2026
- Subject: Computer Vision and Pattern Recognition (cs.CV)
- Length: 13 pages
- Cite as: arXiv:2609.26809 [cs.CV]
- DOI: https://doi.org/10.48550/arXiv.2609.26809
via ArXiv CV